An efficient indexing scheme for image storage and recognition

This paper presents a model-based vision system to achieve robust recognition of planar contours that are scale invariant of known models. Planar contours are partitioned into segments by using constant curvature criterion. A set of descriptors that are invariant with respect to scale, rotation, and translation are extracted from the geometric features of the segments. The descriptors are used to carry out an efficient indexed search over the models so as to reduce the search space. Fragments of contours extracted from partially occluded scenes can be individually matched by using the local shape descriptors. Pruning of large portions of the models is carried out by keeping only some matched classes which received the highest vote. This significantly reduces the search and enables the use of finer matching operators, such as comparing the positioning of segments in the scene to positioning of matched segments in the model. More sophisticated matching is applied in later stages over a much restricted number of hypotheses. Therefore, the dependency of the recognition time over the size of the models is significantly reduced. Evaluation shows the ability of our approach to recognize scenes with real partially occluded objects. Entirely visible objects are recognized with a reasonably high efficiency (80%), even with a change in viewpoint of up to 25°. The efficiency smoothly decreases, but remains above 60% when the percentage of visible segments drops to 50% and the change in viewpoint is as above.